The liblinear java port until 1.7 uses the Java BLAS (Basic Linear
Algebra Subprogram) port from http://www.netlib.org/lapack that is based
on Fortran code.
However, this dependency is an overkill since only rather primitive
functions are needed.

Some unreachable code is dropped [1] and complexity is minimized [2].
Micro benchmarking shows that the new code is even faster while yielding
the same results.
Macro benchmarks show however, that this optimization has almost no
effect on training time.

[1] "incx" and "incy" is always 1
[2] folded the (unrolled) loops. modern compilers are smarter and thus
the execution is faster.